MelodyDeterminism: A Canonical Framework for Bitwise Reproducibility in LLM Inference

Abstract Reproducibility is a key requirement in scientific research, regulatedindustries, and AI safety. Yet, large language models (LLMs) executed on modernaccelerators exhibit non-determinism due to parallel floating-point reductions,scheduling variability, and stochastic primitives. Current mitigations—such asdeterministic kernels or batch-invariant operators—address fragments of theproblem but lack a unifying framework. We propose MelodyDeterminism, analgorithmic approach where every numerical operator is treated as a melody:identified by a fingerprint (“intro”), mapped to a canonical rule (“score”), andexecuted in lockstep by GPU cores (“orchestra”). This ensures that identicalinputs always produce identical outputs, bitwise reproducible across runs,while minimizing overhead by applying determinism only where necessary.